基于安全两方计算的隐私保护线性回归算法

随着数据安全与隐私泄露事件频发，泄露规模连年加剧，如何保证机器学习中数据和模型参数的隐私引发科学界和工业界的广泛关注。针对本地存储计算资源的有限性及云平台的不可信性所带来的数据隐私问题，基于秘密共享技术提出了一种安全两方计算的隐私保护线性回归算法。利用加法同态加密和加法掩码实现了秘密共享值的乘法计算协议，结合小批量梯度下降算法，最终实现了在两个非共谋的云服务器上的安全线性回归算法。实验结果表明，该方案同时保护了线性回归算法训练及预测阶段中的数据及模型参数，且模型预测性能与在明文域中进行训练的模型相近。

Privacy-Preserving Linear Regression Algorithm Based on Secure Two-Party Computation

With the frequent occurrence of data security and privacy leaks and the increasing scale of leaks year after year, how to ensure the privacy of data and model parameters in machine learning has aroused widespread concern in the scientific and industrial community. Aiming at the data privacy problem caused by the limitation of local storage and computing resources and the untrustworthiness of cloud platforms, this paper proposes a privacy protection linear regression algorithm for secure two-party computation based on secret sharing technology. This paper uses additive homomorphic encryption and additive masks to realize a multiplication calculation protocol of the secret shared values, and combines the mini-batch gradient descent algorithm to realize the secure linear regression algorithm on two non-collusive cloud servers. The experimental results show that the scheme protects the data and model parameters in the training and prediction phases of the linear regression algorithm at the same time, and the model prediction performance is similar to the model trained in the plaintext domain.